Midjourney, until now known purely for its image- and video-generation software, used a San Francisco event to reveal its first hardware product: a full-body ultrasound scanner the company is calling the Midjourney Scanner. CEO David Holz pitched the device as eventually rivaling MRI for diagnostic clarity, while taking a fraction of the time.
The scanner works by lowering a person into a shallow pool ringed with ultrasonic transducers that fire sound waves through the body from every angle; Midjourney's compute stack then reconstructs the returning signal into a three-dimensional map of muscle, bone, fat, and organs in roughly a minute. The hardware runs on dozens of Butterfly Network ultrasound-on-chip modules under a co-development deal that sent Butterfly's own stock sharply higher on the news.
Midjourney plans to open a flagship "spa" in San Francisco's Union Square before the end of 2027, outfitted with ten scanners alongside saunas and cold plunges, and has framed the device as one of eight active hardware and software projects inside the company. Regulatory approval remains the open question — Holz has acknowledged the path to FDA clearance for thousands of diagnostic use cases will take years to walk.
References to an unreleased model called GPT-5.6 have been surfacing in OpenAI's Codex backend logs for weeks, and prediction markets are now pricing a launch sometime in the next several days. None of it is official: OpenAI has not announced a release date, specifications, or even confirmed the name.
The headline rumor is a context window of roughly 1.5 million tokens, a meaningful jump from GPT-5.5's documented one million and a figure that would keep pace with Google's aggressive push on long-context models. For developers, the practical upside shows up in agentic and codebase-scale work — fewer manual chunking tricks, more of a sprawling repository or research session held in a single pass.
OpenAI's chief scientist has reportedly told staff internally that the model represents a meaningful step up from GPT-5.5, though that account comes secondhand rather than from an official statement. Until OpenAI says otherwise, GPT-5.5 remains the model actually serving ChatGPT and Codex traffic.
Databricks has agreed to acquire Panther, an AI-powered security operations center platform, in a deal whose financial terms were not disclosed. It's the company's third security acquisition, and it lands squarely inside Databricks' push toward what it calls a "security lakehouse" — a single data layer meant to replace costly, manual legacy SIEM tools.
Panther brings more than 100 prebuilt integrations across cloud, identity, and SaaS systems, along with detection-as-code workflows that let security teams automate alert triage instead of working tickets by hand. Databricks CEO Ali Ghodsi has argued that AI has shrunk the time attackers need to exploit a vulnerability to the point that traditional, human-paced security operations can no longer keep up.
Anthropic is named as one of Panther's customers in the deal announcement, with its head of defense crediting the platform for a more programmable, engineering-style approach to threat detection — notable given how much of the AI industry's own security posture now depends on tools built by other AI companies.
OpenAI has rolled out Record & Replay, a Codex feature that flips the usual prompt-first workflow: instead of describing a task in words, a person simply performs it once on their Mac while Codex watches, then packages the sequence into an editable, reusable "skill."
OpenAI suggests starting with stable, well-defined workflows — filing an expense report, booking a parking spot, submitting a time-off request — and says Codex captures not just the mechanical clicks but underlying preferences, like naming conventions or default fields, that are easier to show than to explain. Skills can be refined after recording and shared across a team, so one person's demo can become a whole department's automation.
The feature is rolling out to select markets first and remains limited to macOS for now, with OpenAI pointing builders toward its broader plugin system if they need to bundle multiple skills or distribute something more durable across an organization.
For years, families with undiagnosed rare diseases have lived inside a particularly cruel kind of uncertainty: a child with clear symptoms, exhaustive testing, and still no name for what's wrong. A new study out of Boston Children's Hospital's Manton Center, Harvard, and OpenAI gave 376 of those previously unsolved cases a second look, this time with an AI reasoning model in the loop.
Using OpenAI's o3 Deep Research model, researchers fed it de-identified clinical and genomic records and asked it to surface evidence-linked candidate explanations. After expert review, additional lab testing, and clinical confirmation, physicians established new diagnoses in 18 of the cases — an added diagnostic yield of roughly 4.8% on top of everything specialists had already found.
The authors frame this as a research workflow, not a product: the model never diagnosed anyone, and OpenAI says the study is not evidence that clinicians or patients should use its models for medical decision-making. What it does show is a case for periodic reanalysis — some answers only surface once enough new knowledge, or enough fragmented records, get pulled back together.
Retell AI used a self-styled "2026 launch week" to ship five product announcements in five days, the centerpiece being a CRM built directly into the platform so customer records stay in sync with every voice call automatically, without a separate integration step.
The company also introduced live call monitoring and custom dashboards aimed at real-time quality assurance, alongside a "Colloquial Model" and an "Expressive Mode" designed to push its voice agents further from the flat, scripted cadence that still gives away a lot of automated calls.
The releases build on a platform that already markets itself on low latency and natural turn-taking in conversation, and they arrive as voice AI vendors broadly compete to close the gap between "clearly a bot" and "clearly not."